ExternalVendorCode/whisper.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/whisper.cpp.git synced 2025-06-13 12:38:08 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

whisper : use flash attention (#2152)

Browse Source 
* whisper : use flash attention in the encoder

* whisper : add kv_pad

* whisper : remove extra backend instance (huh?)

* whisper : use FA for cross-attention

* whisper : use FA for self-attention

* whisper : simplify encoder FA

* whisper : add flash_attn runtime parameter

* scripts : add bench log

* scripts : add M1 Pro bench log
This commit is contained in:
Georgi Gerganov
2024-05-15 09:38:19 +03:00
committed by GitHub
parent 9d5771ae43
commit 7094ea5e75
13 changed files with 657 additions and 172 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
examples/lsp/lsp.cpp
View File

@ -31,6 +31,7 @@ struct whisper_params {
bool print_special = false;
bool print_energy = false;
bool use_gpu = true;
bool flash_attn = false;
std::string language = "en";
std::string model = "models/ggml-base.en.bin";
@ -74,6 +75,7 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
else if (arg == "-ps" || arg == "--print-special") { params.print_special = true; }
else if (arg == "-pe" || arg == "--print-energy") { params.print_energy = true; }
else if (arg == "-ng" || arg == "--no-gpu") { params.use_gpu = false; }
else if (arg == "-fa" || arg == "--flash-attn") { params.flash_attn = true; }
else if (arg == "-l" || arg == "--language") { params.language = argv[++i]; }
else if (arg == "-m" || arg == "--model") { params.model = argv[++i]; }
else {
@ -105,6 +107,7 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
fprintf(stderr, " -ps, --print-special [%-7s] print special tokens\n", params.print_special ? "true" : "false");
fprintf(stderr, " -pe, --print-energy [%-7s] print sound energy (for debugging)\n", params.print_energy ? "true" : "false");
fprintf(stderr, " -ng, --no-gpu [%-7s] disable GPU\n", params.use_gpu ? "false" : "true");
fprintf(stderr, " -fa, --flash-attn [%-7s] flash attention\n", params.flash_attn ? "true" : "false");
fprintf(stderr, " -l LANG, --language LANG [%-7s] spoken language\n", params.language.c_str());
fprintf(stderr, " -m FNAME, --model FNAME [%-7s] model path\n", params.model.c_str());
fprintf(stderr, "\n");
@ -436,7 +439,10 @@ int main(int argc, char ** argv) {
// whisper init
struct whisper_context_params cparams = whisper_context_default_params();
cparams.use_gpu = params.use_gpu;
cparams.use_gpu = params.use_gpu;
cparams.flash_attn = params.flash_attn;
struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
// init audio